基于改进RT-DETR的交通场景图像长尾目标检测方法

doi:10.16182/j.issn1004731x.joss.26-0200

摘要/Abstract

摘要：

检测交通环境中的稀有目标对于自动驾驶系统的安全性和可靠性至关重要。针对现有方法对于训练中仅出现极少次数的稀有目标极端情况检测能力有限的问题，提出基于RT-DETR的交通场景图像中长尾目标检测方法TF-DETR (tail free DETR)。构建轻量化的CSP-EDA (CSP-efficient depthwise aggregation)用以重构骨干网络，其核心单元高效深度聚合模块EDA结合池化-转置注意力与卷积门控单元，实现了对尾部样本特征的增强。在编码器中设计空间自适应动态认知交互模块，通过两阶段之间的特征复用，建立针对尾部目标从特征提取到多尺度认知增强的高效信息交互机制。引入风车卷积对特征图进行下采样，通过非对称填充与多方向卷积核的协同，在轻量化设计的基础上增强对尾部目标的检测能力。在CODA2022数据集上的实验结果表明，TF-DETR相较于基线模型AP50实现了4.9%的提升，AR50提升了5.4%，计算量减少了22.9%，参数量减少了25.7%。

关键词: 交通场景图像, 目标检测, 深度学习, 长尾分布, RT-DETR

Abstract:

Detecting rare objects in transportation environments is crucial for the safety and reliability of autonomous driving systems. To address the challenges of limited detection capability on rare objects that appear only a few times during training, we propose an improved long-tailed object detection method for traffic scene images named TF-DETR (tail free DETR) based on RT-DETR model. First, the lightweight CSP-EDA (CSP-efficient depthwise aggregation) was constructed toreconstruct the backbone network. Its core unit, the efficient depthwise aggregation, integrated the pooled-transpose attention and convolutional gated linear unit to enhance feature representations for tail-class samples. Second,the spatially-adaptive dynamic cognitive interaction was deployed in the encoder. Through feature reuse between the two-stage, the interaction module established an efficient information interaction mechanism. The mechanism spanned from feature extraction to multi-scale cognitive enhancement for tail-class objects. Finally, the pinwheel-shaped convolution wasintroduced for feature map downsampling. Through its collaborative design of asymmetric padding and multi-directional kernels, the model enhanced detection capability for tail-class objects while maintaining a lightweight architecture. Experimental results on the CODA2022 dataset show that TF-DETR achieves a 4.9% increase in AP50 and a 5.4% increase in AR50 compared with the baseline. Meanwhile, a 22.9% reduction in FLOPs and a 25.7% decrease in Parameters are realized.

Key words: transportation scene images, object detection, deep learning, long-tail distribution, RT-DETR

中图分类号:

U495

邓明君,聂沛华,郭军华等 . 基于改进RT-DETR的交通场景图像长尾目标检测方法[J]. 系统仿真学报, 2026, 38(9): 2518-2534.

Deng Mingjun,Nie Peihua,Guo Junhua,et al . Long-tail Object Detection Method in Transportation Scene Images Based on improved RT-DETR[J]. Journal of System Simulation, 2026, 38(9): 2518-2534.

图/表 16

图1

图2

图3

图4

图5

图6

表1

表2

表3

表4

表5

图7

图8

图9

图10

图11

参考文献 27

[1]	Breitenstein Jasmin, Termöhlen Jan-Aike, Lipinski Daniel, et al. Systematization of Corner Cases for Visual Perception in Automated Driving[C]//2020 IEEE Intelligent Vehicles Symposium (IV). Piscataway: IEEE, 2020: 1257-1264.
[2]	Girshick R, Donahue J, Darrell T, et al. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2014: 580-587.
[3]	Liu Wei, Anguelov D, Erhan D, et al. SSD: Single Shot MultiBox Detector[C]//Computer Vision — ECCV 2016. Cham: Springer International Publishing, 2016: 21-37.
[4]	Redmon J, Divvala S, Girshick R, et al. You Only Look Once: Unified, Real-time Object Detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2016: 779-788.
[5]	Carion N, Massa F, Synnaeve G, et al. End-to-end Object Detection with Transformers[C]//Computer Vision — ECCV 2020. Cham: Springer International Publishing, 2020: 213-229.
[6]	Zhao Yian, Wenyu Lü, Xu Shangliang, et al. DETRs Beat YOLOs on Real-time Object Detection[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 16965-16974.
[7]	Gao Erfeng, Huang Weiguo, Shi Juanjuan, et al. Long-tailed Traffic Sign Detection Using Attentive Fusion and Hierarchical Group Softmax[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(12): 24105-24115.
[8]	邓召学, 刘灿, 郝丙森, 等. 面向长尾分布和特征缺失的交通标志目标检测算法[J]. 吉林大学学报(工学版), 2026, 56(6): 1642-1654.
	Deng Zhaoxue, Liu Can, Hao Bingsen, et al. Object Detection Algorithm for Traffic Signs in Long-tail Distribution and Feature Incompleteness[J]. Journal of Jilin University(Engineering and Technology Edition), 2026, 56(6): 1642-1654.
[9]	伍云鹏, 付应雄, 沈丽君, 等. 基于YOLOX-tiny的长尾分布交通标志识别模型[J]. 系统仿真学报, 2024, 36(11): 2503-2516.
	Wu Yunpeng, Fu Yingxiong, Shen Lijun, et al. Traffic Sign Recognition Model with Long-tail Distribution Based on YOLOX-tiny[J]. Journal of System Simulation, 2024, 36(11): 2503-2516.
[10]	Xiong Zhongxia, Song Tao, He Shan, et al. A Unified and Costless Approach for Improving Small and Long-tail Object Detection in Aerial Images of Traffic Scenarios[J]. Applied Intelligence, 2023, 53(11): 14426-14447.
[11]	Zhang Hancheng, Hu Yuanyuan, Qian Zhendong, et al. Enhancing Rare Object Detection on Roadways Through Conditional Diffusion Models for Data Augmentation[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(11): 19018-19029.
[12]	Zhang Houwang, Leanne Lai Hang Chan. Overcoming the Challenges of Long-tail Distribution in Nighttime Vehicle Detection[J]. IEEE Intelligent Systems, 2024, 39(2): 51-60.
[13]	Wang Chien-Yao, Liao Hongyuan, Wu Yueh-Hua, et al. CSPNet: A New Backbone that Can Enhance Learning Capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Piscataway: IEEE, 2020: 1571-1580.
[14]	Noman Mubashir, Fiaz Mustansar, Cholakkal Hisham, et al. ELGC-net: Efficient Local-global Context Aggregation for Remote Sensing Change Detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 1-11.
[15]	Shi Dai. TransNeXt: Robust Foveal Visual Perception for Vision Transformers[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 17773-17783.
[16]	Zhou Shihao, Chen Duosheng, Pan Jinshan, et al. Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 2952-2963.
[17]	Zhu Jiachen, Chen Xinlei, He Kaiming, et al. Transformers Without Normalization[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2025: 14901-14911.
[18]	Yin Dongshuo, Hu Leiyi, Li Bin, et al. 5%>100%: Breaking Performance Shackles of Full Fine-tuning on Visual Recognition Tasks[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2025: 20071-20081.
[19]	Chen Shuang, Zhang Haozheng, Atapour-Abarghouei A, et al. SEM-net: Efficient Pixel Modelling for Image Inpainting with Spatially Enhanced SSM[C]//2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Piscataway: IEEE, 2025: 461-471.
[20]	Yang Jiangnan, Liu Shuangli, Wu Jingjun, et al. Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection[C]//Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence. Palo Alto: AAAI Press, 2025: 9202-9210.
[21]	Li Kaican, Chen Kai, Wang Haoyu, et al. CODA: A Real-world Road Corner Case Dataset for Object Detection in Autonomous Driving[C]//Computer Vision — ECCV 2022. Cham: Springer Nature Switzerland, 2022: 406-423.
[22]	Chengqi Lü, Zhang Wenwei, Huang Haian, et al. RTMDet: An Empirical Study of Designing Real-time Object Detectors[EB/OL]. (2022-12-16) [2026-02-25]. .
[23]	Zhang Shifeng, Chi Cheng, Yao Yongqiang, et al. Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 9756-9765.
[24]	Wang Chengcheng, He Wei, Nie Ying, et al. Gold-YOLO: Efficient Object Detector via Gather-and-distribute Mechanism[C]//Proceedings of the 37th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2023: 51094-51112.
[25]	Wang Chien-Yao, Yeh I-Hau, Liao Hongyuan. YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information[C]//Computer Vision — ECCV 2024. Cham: Springer Nature Switzerland, 2025: 1-21.
[26]	Wang Ao, Chen Hui, Liu Lihao, et al. YOLOv10: Real-time End-to-end Object Detection[C]//Proceedings of the 38th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2024: 107984-108011.
[27]	Tian Yunjie, Ye Qixiang, Doermann D. YOLOv12: Attention-centric Real-time Object Detectors[EB/OL]. (2025-02-18) [2026-02-25]. .

实验	CSP-EDA	SDCI	PSConv	AP50/%	AP75/%	AR50/%	AR75/%	FLOPs	Parameters
1				44.0	34.8	53.2	42.5	57.1×10⁹	19.92×10⁶
2	√			45.6	35.2	57.1	42.9	43.1×10⁹	12.87×10⁶
3		√		44.3	35.7	54.6	43.6	58.5×10⁹	22.30×10⁶
4			√	44.4	34.9	54.6	43.8	56.6×10⁹	19.46×10⁶
5	√	√		46.9	36.1	58.4	45.5	44.5×10⁹	15.25×10⁶
6	√		√	46.3	36.2	57.3	44.3	42.6×10⁹	12.41×10⁶
7		√	√	45.4	35.6	56.7	43.7	58.0×10⁹	21.84×10⁶
8	√	√	√	48.9	38.6	58.6	46.6	44.0×10⁹	14.79×10⁶

Model	AP50/%	AP75/%	AR50/%	AR75/%	FLOPs	Parameters
RT-DETR	44.0	34.8	53.2	42.5	57.1×10⁹	19.92×10⁶
+Mambaout	41.8	32.5	51.9	40.3	42.2×10⁹	16.00×10⁶
+Starnet	37.8	28.4	50.1	36.5	40.3×10⁹	15.84×10⁶
+Lsknet	42.1	33.1	53.9	41.7	59.6×10⁹	20.47×10⁶
+Mobilenetv4	40.5	31.4	51.0	38.4	39.7×10⁹	11.36×10⁶
+Repvit	42.5	32.6	54.5	41.5	60.5×10⁹	22.26×10⁶
+Convnextv2	38.5	29.3	48.9	36.0	63.5×10⁹	24.41×10⁶
+Efficientvit	42.2	32.6	53.0	40.1	52.8×10⁹	20.81×10⁶
+CSP-EDA (ours)	45.6	35.2	57.1	42.9	43.1×10⁹	12.87×10⁶

Model	AP50/%	AP75/%	AR50/%	AR75/%	FLOPs	Parameters
RT-DETR+CSP-EDA	45.6	35.2	57.1	42.9	43.1×10⁹	12.87×10⁶
+EDFFN	42.2	33.1	53.5	41.4	43.0×10⁹	12.76×10⁶
+SEFFN	44.0	35.3	52.6	42.0	43.0×10⁹	12.75×10⁶
+SEFN	45.4	34.8	55.0	42.4	43.6×10⁹	14.33×10⁶
+DyT	43.9	34.2	54.6	42.3	43.1×10⁹	12.87×10⁶
+ASSA	45.1	35.1	55.0	43.9	44.0×10⁹	13.71×10⁶
+Mona	46.3	36.5	57.7	45.0	43.2×10⁹	12.96×10⁶
+SDCI (ours)	46.9	36.1	58.5	45.6	44.5×10⁹	15.25×10⁶

Model	AP50/%	AP75/%	AR50/%	AR75/%	FLOPs	Parameters
RT-DETR+CSP-EDA+SDCI	46.9	36.1	58.5	45.6	44.5×10⁹	15.25×10⁶
+Wconv	44.6	35.7	55.6	43.5	42.1×10⁹	15.25×10⁶
+FourierConv	43.6	32.8	56.2	41.0	44.5×10⁹	17.36×10⁶
+GSConvE	44.8	34.0	56.1	43.1	43.9×10⁹	14.96×10⁶
+SRFD	44.9	35.3	55.1	43.5	44.0×10⁹	15.00×10⁶
+HWD	42.7	33.1	53.3	40.2	43.2×10⁹	14.60×10⁶
+PSConv (ours)	48.9	38.6	58.6	46.6	44.0×10⁹	14.79×10⁶

Model	AP50/%	AP75/%	AR50/%	AR75/%	FLOPs	Parameters
RT-DETR	44.0	34.8	53.2	42.5	57.1×10⁹	19.92×10⁶
RTMDet^[22]	29.0	21.4	53.7	34.5	39.1×10⁹	24.68×10⁶
SSD^[3]	27.5	16.3	38.5	20.3	32.5×10⁹	29.36×10⁶
ATSS^[23]	31.0	23.0	58.0	38.9	110.0×10⁹	38.91×10⁶
Gold-YOLO^[24]	35.5	25.4	58.3	42.7	87.3×10⁹	41.30×10⁶
YOLOv8m	40.9	30.6	62.6	44.5	78.8×10⁹	25.86×10⁶
YOLOv9m^[25]	41.9	30.9	63.4	41.3	76.7×10⁹	20.04×10⁶
YOLOv10m^[26]	38.4	29.0	59.9	44.2	59.0×10⁹	15.33×10⁶
YOLO11m	40.9	30.9	61.0	45.3	67.8×10⁹	20.06×10⁶
YOLO12m^[27]	41.1	31.4	58.5	40.9	67.3×10⁹	20.10×10⁶
TF-DETR (ours)	48.9	38.6	58.6	46.6	44.0×10⁹	14.79×10⁶